Analyzing issues with PZ (4/4)

[Klazkin]

Explain the context of the issue, what is being addressed in detail.

The goal

• [x] Higher epoch and datasamples training.
• [x] GPS model.
• [x] adjustments to GPS block.
• [ ] Investigate what causes the policy training to not go above 50% accuracy (is it a model architecture limitation?)
• [x] Verify MCTS score calucalted correctly.
• [x] Minor MCTS algorithm adjustments.
• [x] Increased model base specs
• [x] #97
  • During self-play, instead of chosing max action, they are instead chosen using a temperature metric -> higher temp means choices are more random, lower temp -> more often picks high score actions
  • [x] Analyze the impact of using a weighted scoring function instead of mean in pure MCTS
  • [x] ignore training samples with only 1 action choice
  • [x] check handling of empty samples too

Time tracking

Time Estimate: 0 hours 0 minutes Time spent: 4 hours 20 minutes

Resources

...